The scenario that is happening more and more often
You get a call. It's your son, in a panic: "Mum, I've had an accident, you need to send me 5,000 euros right now, I can't talk for long." His voice is shaking, but it is his voice. You recognise his intonation, his pace, his verbal tics. Except it isn't your son. It's an AI that has cloned his voice from a few seconds of audio taken from an Instagram story posted two years ago.
This scenario, which sounded like science fiction in 2022, had by 2026 become one of the most widespread scams in the world. According to the "State of Voice-Based Fraud 2026" report, 84% of financial and retail organisations have faced sophisticated voice attacks over the past year. And according to Deloitte (one of the leading global audit and consulting firms), losses linked to AI fraud in the United States will reach 40 billion dollars by 2027.
This article takes a precise look at this threat, the tools used, the documented real-world cases, and what you can concretely do to protect yourself.
3 seconds of audio: what is enough today
The technology crossed a critical threshold in 2024-2025. Where three years ago it took 20 to 30 minutes of audio recording to clone a voice properly, 3 to 10 seconds are now enough with the commercial tools publicly available. The main tools:
- ElevenLabs: a British company founded in 2022, valued at more than 3 billion dollars. Probably the world leader in voice cloning and speech synthesis. Offers thousands of voices in dozens of languages. As we documented in our article on the AI startups that are winning, it is one of the most dynamic success stories in the sector.
- LOVO: more than 500 voices in 100 languages, mainly used for dubbing and content creation.
- Speechify: more than 1,000 voices in 60 languages, founded in 2017, initially to help people with dyslexia.
- VEED: more than 35 voices able to speak in dozens of languages, focused on video creation.
- Tortoise TTS and XTTS: open source alternatives (freely accessible software), technically just as capable as commercial solutions, but without the ethical safeguards that companies put in place.
The psychological threshold has now been crossed: according to several studies, humans correctly identify a deepfake voice only slightly better than chance. On a phone call, where audio quality is degraded anyway, detection becomes practically impossible.
On 16 April 2026, US Senator Maggie Hassan (Democratic Senator for New Hampshire, vice-chair of the Joint Economic Committee, Congress's joint economic commission) sent official letters to ElevenLabs, LOVO, Speechify and VEED, demanding that they explain how they prevent criminal exploitation of their services. She asks them to set out whether they monitor fraudulent use, verify consent before cloning, detect attempts to impersonate public figures or minors, watermark their outputs (cryptographic marking), and report malicious actors to the authorities. This intervention by the US Congress marks the entry of the issue onto the top-tier political agenda.
The real cases documented in 2024-2026
To understand the scale of the phenomenon, let us examine the most emblematic cases.
Hong Kong, February 2024: 25 million dollars gone. An employee in the financial department of a multinational took part in a video conference with his CFO (Chief Financial Officer) and several other company executives. All of them asked him to urgently authorise a transfer of 25 million dollars to specific accounts. The employee complied. Except that every person present on the video conference was a deepfake generated by AI. The real CFO was never aware of it. It is the largest deepfake theft ever documented to date.
United Kingdom, 2023: 220,000 euros via a cloned voice. A British energy company lost around 220,000 euros following a phone call imitating the voice of the CEO of the German parent company. The voice, with an identical German accent, requested an urgent transfer to a Hungarian supplier. The call was entirely fabricated by AI.
Arizona, 2023: fake kidnapping. A mother received a call imitating the voice of her crying daughter, claiming she had been kidnapped. The "kidnappers" demanded a ransom. The daughter was in fact at her school, safe. Several similar cases were documented by US police in 2024-2025.
New Hampshire, 2025: grandparent scam campaign. According to an investigation cited by Senator Hassan, several families in New Hampshire were targeted by calls imitating the voice of their child or grandchild claiming to have been arrested and asking for money for bail. One victim describes having been "convinced because the story was delivered in my son's voice, full of terror".
Suffolk County, early 2025: seniors targeted en masse. According to the FBI, victims over 60 lost 4.9 billion dollars to cybercrime in 2024, a rise of 43% compared with 2023. AI is now identified as a major accelerator of these scams. Hiya, a company specialising in detecting fraudulent calls, found that a third of those surveyed across several countries had encountered deepfake voice fraud, with an average loss of 6,000 dollars per victim.
The anatomy of a voice attack in 4 steps
To understand how to protect yourself, you need to understand how these attacks actually work. Group-IB (a Singaporean cybersecurity firm) published a detailed analysis of the full process in 2025, in four stages.
Stage 1: collecting the target's audio. The scammers gather a few seconds of their target's voice. The main sources: social media (Instagram, TikTok and Snapchat videos), online conferences, podcasts, hacked WhatsApp voice messages, old recorded calls. A few seconds publicly available on the internet are enough. For a company executive who speaks in public often, the raw material is endless.
Stage 2: cloning with AI. The audio file is run through a cloning tool (ElevenLabs, LOVO, or open-source alternatives). The result is a voice model capable of producing any text in the target's voice. The training phase takes only a few minutes.
Stage 3: setting the scene. The scammers create an emergency scenario. For individuals: a fake kidnapping, an accident, an arrest. For companies: an urgent transfer requested by the CEO, rapid approval of an invoice, authorisation of a supplier payment. The scenario has to create enough emotional pressure to bypass rational verification.
Stage 4: the call and the transfer. The call is made, often with caller ID spoofing (a technique that makes a fake number appear on the recipient's screen, for example the real CEO's or a family member's). If the victim takes the bait, the transfer goes through. The funds are then quickly laundered through multiple bank accounts, crypto wallets, or shell companies.
Caller ID spoofing combined with voice cloning makes these frauds extremely convincing. When your phone displays "Mum" or "CEO's office" while you hear the matching voice, your brain switches into "known person" mode and automatically disables most of your critical defences. It is this double deception (identifier + voice) that explains why executives who have been trained in cybersecurity nonetheless fall into the trap.
Why your banks are panicking
For the banking sector, voice cloning is an operational nightmare. For years, several banks (HSBC, JP Morgan, Barclays, Crédit Mutuel in France) had invested heavily in voice biometrics (technology that identifies a person by their voice, considered as reliable as fingerprints). It was presented as the future of secure authentication. All of that is now collapsing.
Several studies in 2025 showed that voice biometric systems can be bypassed by AI-cloned voices in more than 50% of cases. Banks are gradually abandoning these systems or supplementing them with other mechanisms. In the United States, several major banks have already announced that they are dropping voice biometrics as a sole authentication method.
The Experian report "Future of Fraud Forecast 2026" describes the current moment as a "tipping point", with 72% of business leaders identifying AI-assisted fraud as a major operational challenge. Signicat, another cybersecurity player, reports that deepfake fraud attempts have risen by more than 2,000% in three years.
| Indicator | 2026 figure |
|---|---|
| Audio needed to clone a voice | 3 to 10 seconds |
| Rise in voice fraud, late 2024 | +442% |
| Rise in deepfake attempts over 3 years | +2,000% |
| Financial organisations attacked in 2026 | 84% |
| Losses among US seniors, 2024 | $4.9 billion |
| Average loss per voice fraud | $5,000 to $25,000 |
| Projected US AI fraud losses, 2027 | $40 billion |
Seven practical steps to protect yourself
No panic, no paranoia. Just seven habits that make a real difference.
1. Set up a family code word. Agree with your close family on a secret word known only to you. If someone calls claiming to be your child in an emergency, ask for that code word. No scammer will know it, whatever the quality of the voice cloning. It is probably the most effective defence in 2026.
2. Never decide under pressure. Fraudsters always create artificial urgency. The simple rule: "if someone asks you for money urgently over the phone, hang up, call back on the usual number, verify." A genuine family emergency never requires a decision in 30 seconds without verification.
3. For businesses, introduce a dual-authorisation procedure. Any transfer above a certain threshold should require authorisation via a different channel: email + phone, or phone + signature in person. That is what would have saved the Hong Kong company from the $25 million loss.
4. Be wary of probing questions. If a call asks you unusual questions ("do you remember our holiday in Marseille in 2019?"), it is probably to harvest verification details. Give as little personal information as possible to unverified telephone contacts.
5. Limit your public voice exposure. If you are an executive or a public figure, be aware that any public audio of you (interviews, podcasts, videos) can be used to clone your voice. That is not a reason to disappear from the media, but it is a reason to have strict internal protocols in your company.
6. Enable bank alerts. Most banks now allow you to enable immediate notifications for any transfer above a threshold. A notification received within seconds can make it possible to cancel a fraudulent transfer before it becomes final. In France, the legal deadline for contesting a fraudulent SEPA transfer is 13 months, but effective cancellation depends on how quickly you react.
7. Always report it. In France, 33 700 lets you report fraudulent texts and calls. The PHAROS service (Plateforme d'harmonisation, d'analyse, de recoupement et d'orientation des signalements, run by the national police) collects reports of cybercrime. The more reports the authorities have, the better they can identify the networks.
What companies do (or should do)
Several defensive technologies are emerging. None is a miracle cure, but their combination helps.
Audio watermarking. Several players (Google, ElevenLabs, Microsoft) now embed invisible cryptographic signatures in their AI-generated audio. In theory, a detector can identify that an audio clip was generated. In practice, these watermarks can be circumvented, and open-source tools do not include them.
The C2PA standard. The Coalition for Content Provenance and Authenticity (bringing together Adobe, Microsoft, Google, the BBC and several others) is developing a provenance standard for digital content, audio included. As we explored in our article on deepfakes, this standard could eventually reverse the logic: proving that content is authentic rather than proving that it is fake.
"Verified Human" badges in telephony. Several telecoms operators are working on badges displayed on incoming calls to certify that the caller is a verified human and not an AI. Progressive rollout planned for 2026-2028.
Real-time AI detection. Several startups (Reality Defender, Pindrop, Hiya) are developing systems capable of detecting a synthetic voice in real time during a call. But only 22% of financial institutions have deployed AI fraud prevention tools, which creates a dangerous gap: attackers are using AI, but defences remain largely manual.
The regulation that is coming (slowly)
Several regulatory frameworks are emerging.
United States. The FCC (Federal Communications Commission, the US federal agency that regulates telecommunications) officially declared robocalls (automated mass calls) using AI-generated voices illegal in February 2024. Several states (Tennessee, California) have passed their own stricter laws. Congress is currently examining several federal bills specifically targeting voice fraud.
European Union. The AI Act imposes transparency obligations on all AI-generated content. Several member states are working on specific legislation on voice deepfakes. The UK and the EU are working on mandates for banks to reimburse victims of APP fraud (Authorized Push Payment fraud: fraud where the victim is manipulated into authorising the transfer themselves).
France. The law of 21 May 2024 against digital violence covers certain aspects of audio deepfakes. But no law specific to AI voice fraud yet exists.
As we analysed in our article on the limits of AI regulation, these provisions are necessary but structurally limited by the availability of open source tools. Tortoise TTS and XTTS can generate cloned voices with no obligation of watermarking, no transparency, no traceability.
The brutal takeaway to remember
AI voice fraud is not a future threat. It is a daily reality in 2026. The figures show an exponential acceleration that will not slow down for several years, and probably never completely. As we explored in our article on the environmental cost of AI and our analysis of AI ghost workers, every advance in AI creates side effects that society must learn to manage.
>>The good news: simple defences (code words, two-factor verification, bank alerts) are effective against 90% of attacks. The bad news: you have to put them in place actively, because nobody will do it for you. While regulation catches up with the technology, it is up to individuals and businesses to adapt. To go further, see also our article on faked AI demos, which explores the same kind of perceptual manipulation in a different context.
One last thing. If you are a parent, take 10 minutes this evening to agree on a code word with your children. It is not paranoid. It has simply become necessary in 2026.